A #FoundationModel is just the first step in revolutionizing #AI for #pathology. But the big question remains: How can we effectively aggregate thousands of tile embeddings for a whole slide? Here's the answer, hot off the press by my exceptional PhD student Shengjia Chen and team: 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠 𝐀𝐠𝐠𝐫𝐞𝐠𝐚𝐭𝐢𝐨𝐧 𝐌𝐞𝐭𝐡𝐨𝐝𝐬 𝐢𝐧 𝐂𝐨𝐦𝐩𝐮𝐭𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐏𝐚𝐭𝐡𝐨𝐥𝐨𝐠𝐲: 𝐀 𝐂𝐥𝐢𝐧𝐢𝐜𝐚𝐥 𝐃𝐚𝐭𝐚 𝐏𝐞𝐫𝐬𝐩𝐞𝐜𝐭𝐢𝐯𝐞 arXiv: https://lnkd.in/dUnx6QqT 🔍 Key Insights: 1️⃣ 𝐃𝐨𝐦𝐚𝐢𝐧-𝐒𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐌𝐨𝐝𝐞𝐥𝐬: We found that embeddings derived from histological image-trained foundation models outperform those based on ImageNet models. 2️⃣ 𝐒𝐩𝐚𝐭𝐢𝐚𝐥-𝐀𝐰𝐚𝐫𝐞 𝐀𝐠𝐠𝐫𝐞𝐠𝐚𝐭𝐢𝐨𝐧: While spatial-aware aggregators enhance performance with ImageNet models, they offer less improvement when using pathology foundation models. 3️⃣ 𝐍𝐨 𝐎𝐧𝐞-𝐒𝐢𝐳𝐞-𝐅𝐢𝐭𝐬-𝐀𝐥𝐥: No single aggregation method excels across all tasks, highlighting the need for adaptable techniques in clinical AI. 👉️ 𝐁𝐨𝐧𝐮𝐬: Checkout the research landscape map to see the evolution of aggregation methods over time! This work took a village, I am deeply grateful to have been part of: Shengjia Chen, Gabriele Campanella, Abdulkadir Elmas, Aryeh Stock, Jennifer Z., Alexandros D. Polydorides, Adam Schoenfeld, Kuan-lin Huang, Jane Houldsworth, Chad Vanderbilt #computationalpathology #histology #computervision #machinelearning #biotech #healthcare #deeplearning
Thanks for sharing
Congratulations Shengjia Chen !! Very cool work!
Thanks for sharing! Great
Interesting! Thanks for sharing.
Great!
Senior Director-Digital Pathology Leader-Science Fellow/Oncology Precision Medicine at Bayer Pharmaceuticals
1moNice! Thanks for sharing Thomas Fuchs !